Assessing the Relationship between Economic News Coverage and Mass Economic Attitudes

Acknowledgments
We thank Jamie Monogan, Markus Prior, and Walt Stone for 
helpful comments on previous drafts. We are also immensely 
grateful to Pablo Barberá, Ryan McMahon, Jonathan Nagler, 
Stuart Soroka, Dominik Stecula, and Christopher Wlezien for 
their general support and insights on the relationship between 
economic conditions, media coverage, and mass attitudes.
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with 
respect to the research, authorship, and/or publication of this 
article.
Funding
The author(s) received no financial support for the research, 
authorship, and/or publication of this article.
Notes
1. We will measure this “extra-economic” coverage with the 
residuals from a model of economic news coverage purged 
of the influence of economic performance.
2. Economic performance may also influence the tone of eco-
nomic media coverage, a relationship we consider below, 
which further complicates efforts to assess the media/con-
sumer sentiment relationship.
3. This is not to say the media has no effect. Rather, a strong 
connection between reality (national economic perfor-
mance) and opinion (collective economic assessments) 
could be produced without the media playing a causal role.
4. The leading economic indicator index as provided by the 
Conference Board is a weighted average of: (1) average 
weekly hours (manufacturing), (2) average weekly initial 
claims for unemployment insurance, (3) manufacturers’ 
new orders (consumer goods and materials), (4) the ISM® 
Index of New Orders, (5) manufacturers’ new orders (non-
defense capital goods excluding aircraft orders), (6) build-
ing permits (new private housing units), (7) stock prices 
(five hundred common stocks), (8) leading credit index, 
(9) interest rate spread (ten-year treasury bonds, less fed-
eral funds), and (10) average consumer expectations for 
business conditions. Soroka, Stecula, and Wlezien (2015) 
have a version of the index excluding consumer expecta-
tions that are included in the Index of Consumer Sentiment 
(ICS) from the University of Michigan consumer surveys.
5. The index of lagging economic indicators includes: (1) the 
average duration of unemployment (measured in weeks, 
sign inverted), (2) inventory to sales ratio (manufacturing 
and trade), (3) labor costs per unit of output (manufactur-
ing), (4) average prime rate charged by banks, (5) volume 
of business loans held by banks and commercial paper 
issued by nonfinancial companies, (6) consumer install-
ment credit to personal income ratio, and (7) the consumer 
price index for services.
6. The index of coincident economic indicators includes: (1) 
payroll employment—changes represent the net hiring and 
firing of nonagricultural businesses, (2) personal income 
(less transfer payments, inflation adjusted), (3) industrial 
production index (covers physical output of all stages of 
production in manufacturing, mining, gas, and electric util-
ity industries), and (4) real manufacturing and trade sales.
7. We omit the Conference Board Index of Leading Economic 
Indicators because the index includes one of the compo-
nents of the ICS.
8. https://www.conference-board.org/data/bci/index.
cfm?id=2160 (accessed January 27, 2017).
9. http://www.econ.yale.edu/~shiller/data.htm (accessed 
January 27, 2017).
10. In deciding how many covariates to include, we need to 
weigh the drawbacks of over-fitting versus those of under-
fitting. If our primary goal was to interpret the coefficients 
of the economic indicators, then including so many covari-
ates would be problematic. But because our primary goal 
was to purge each measure of all possible direct influences 
of economic performance, we were more concerned with 
under-fitting than over-fitting, thus making it appropriate 
to include so many variables.
11. Barberá et al. (2016) compare the predictive accuracy 
(relative to human coding) of measures of tone of the U.S. 
economy as presented in the New York Times 1980–2011 
generated by supervised machine learning (SML) and by a 
number of sentiment dictionaries, including the Lexicoder 
Sentiment Dictionary (Young and Soroka 2012) used 
by Soroka, Stecula, and Wlezien (2015), Sentistrength 
(Thelwall et al. 2010), and Hopkins and King’s (2010) 
nine-word index, which consists of counting the number 
of articles per month mentioning nine economic words 
(inflat, recess, unempl, slump, layoff, jobless, invest, grow, 
growth). They find the accuracy of SML classifications 
to be significantly greater (between 71% and 74%) than 
Lexicoder (
≈57%), Sentistrength (≈57%), and Hopkins 
and King’s index (39%) and that their measures of eco-
nomic tone are more highly correlated with the set of eco-
nomic indicators included here.
12. Using Proquest, Barberá et al. (2016) downloaded all arti-
cles in the four newspapers that contained any of the fol-
lowing terms in parentheses (employment, unemployment, 
inflation, consumer price index, GDP, gross domestic prod-
uct, interest rates, household income, per capita income, 
stock market, federal reserve, consumer sentiment, reces-
sion, economic crisis, economic recovery, globalization, 
outsourcing, trade deficit, consumer spending, full employ-
ment, average wage, federal deficit, budget deficit, gas 
price, price of gas, deflation, existing home sales, new home 
sales, productivity, retail trade figures, wholesale prices) and 
the term United States. This search produced 108,186 sto-
ries from the New York Times (January 1947 to December 
2014), 62,213 from the Washington Post (January 1951 
to December 2014), 69,018 from the Wall Street Journal 
(January 1984 to December 2014), and 21,961 from USA 
Today (April 1957 to December 2014). After download-
ing the articles, they removed any article that mentioned 
any country name, country capital, nationality, or continent 
name (Schrodt 2011) in the headline or first thousand char-
acters of the articles and that did NOT mention U.S., U.S.A., 
or United States in that same text fragment.

---